# Neural network in Flux.jl using CUDA is slower

**URL:** <https://discourse.julialang.org/t/neural-network-in-flux-jl-using-cuda-is-slower/43100>\
**Category:** General Usage\
**Created:** [July 15, 2020, 2:47am UTC](https://discourse.julialang.org/t/neural-network-in-flux-jl-using-cuda-is-slower/43100 "2020-07-15T02:47:13Z")\
**Posts on this page:** 1\
**Page:** 1

<div class="post-metadata">

**Author:** ![xiaodai](https://sea2.discourse-cdn.com/julialang/user_avatar/discourse.julialang.org/xiaodai/32/15937_2.png) [@xiaodai](https://discourse.julialang.org/u/xiaodai)\
**Post date:** [July 15, 2020, 2:47am UTC](https://discourse.julialang.org/t/neural-network-in-flux-jl-using-cuda-is-slower/43100/1 "2020-07-15T02:47:13Z")

</div>

I definied the same neural network on CPU and GPU but the GPU version is slower, and it’s not due to data transfer as I have already `cu`ed the data.

Can anyone shed light on this? I am on Julia 1.5rc1 and Flux 0.11 and CUDA v1.1.0 and my GPU is a reasonable RTX2080

```julia
# in this attempt, I shall play a set of games with the policy and move the policy using
# the data
using Game2048
using Flux, CUDA
CUDA.allowscalar(false)

policy = Chain(
    board->reshape(board, 4, 4, 1, 1),
    Conv((2,2), 1 => 256, relu),
    Conv((2,2), 256 => 128, relu),
    Conv((2,2), 128 => 64, relu),
    flatten,
    Dense(64, 4),
    softmax
) |> gpu

board = initboard()

cb = cu(Float32.(board))

using BenchmarkTools
@benchmark policy(cb)
# BenchmarkTools.Trial:
# memory estimate: 60.33 KiB
# allocs estimate: 1421
# --------------
# minimum time: 418.799 μs (0.00% GC)
# median time: 470.701 μs (0.00% GC)
# mean time: 520.493 μs (2.57% GC)
# maximum time: 36.679 ms (43.01% GC)
# --------------
# samples: 9507
# evals/sample: 1

policy_cpu = Chain(
    board->reshape(board, 4, 4, 1, 1),
    Conv((2,2), 1 => 256, relu),
    Conv((2,2), 256 => 128, relu),
    Conv((2,2), 128 => 64, relu),
    flatten,
    Dense(64, 4),
    softmax
)

board = initboard()
fb = Float32.(board)
@benchmark policy_cpu(fb)
# BenchmarkTools.Trial:
# memory estimate: 153.36 KiB
# allocs estimate: 211
# --------------
# minimum time: 232.399 μs (0.00% GC)
# median time: 295.501 μs (0.00% GC)
# mean time: 310.653 μs (2.14% GC)
# maximum time: 4.700 ms (91.46% GC)
# --------------
# samples: 10000
# evals/sample: 1

```